IRIS: Intelligent Document Histories — Paper Explained

0:00 The vanished draft

Picture Mara hesitating at the door. The finished sentence shows her choice of word, but not how the writer got there.

“Stopped” and “paused” are two alternatives the writer might have considered, though neither appears in the finished sentence.

Watch those alternatives get crossed out, while the finished sentence stays put.

A finished text does not show every decision behind it. So how can writers explore that hidden history?

We’ll explore ordinary writing histories, how IRIS works, and what its studies and limitations tell us.

0:34 From coarse snapshots to process

Here, the three dated drafts show what a conventional history keeps: the text changes in order, from “stopped” to “paused” to “hesitated.”

So the history gives us a chronological record of content changes. But does that show all the writing that led to the final sentence?

Okay, let's turn to the writing traces. Alongside those saved versions, these marks include pauses and temporary wording, giving us a glimpse of activity that a snapshot alone would miss.

And that’s the gap: much writing is temporary, so the final text leaves the decisions and iteration behind it unseen.

That gap leads to the paper’s question, which is how we can design with writing traces to help writers understand their writing process. The snapshots and traces stay in the background, but the question is about making those traces useful for understanding how writing unfolds.

The authors call their tool IRIS, short for In-situ Revision Inquiry System. Its name points to the aim here: bringing writing activity together with document history, rather than treating history as only a sequence of saved texts.

So IRIS joins document history with signals about the writing process, giving writers a way to navigate past versions and reflect on how they wrote. But what those activity signals can actually tell us about a writer’s process is the question we’ll take up next.

2:01 How IRIS makes history explorable

The finished text hides many of the choices that led to it. IRIS starts with the document and a record of typing and editing, including pauses.

Flower and Hayes is a model of writing that groups activity into three kinds, not three stages in a fixed order.

Planning means deciding what to say.

Translating means putting ideas into written words, not translating between languages.

Reviewing means reconsidering what has been written.

The earlier states are arranged chronologically and carry inferred activity labels, linking document changes with the writing activity IRIS attributes to them.

IRIS organizes document states using inferred writing activities. The activity is inferred from the keystroke record, not read directly from the writer’s thoughts.

Those states can give context to a writer’s question. A language model is an AI system that produces responses in words; here, the question enters the model.

Relevant history moves from the earlier states toward the model. This is retrieval augmentation, supplying recorded history as context for its response.

The model receives relevant history when it responds, connecting the writer’s question to earlier document states.

The history is built from traces and inferred activities. How do writers explore it? Figure 1 shows four complementary interface views.

In revision highlighting, hovering over changed text reveals local versions and their differences, so writers can inspect history beside a revision.

Process and topic controls narrow the revision list by a writing activity or a topic the writer chooses.

Process Analysis presents an analysis of earlier writing, while Ask Your Writing lets writers ask questions about their history.

Writers can navigate history through text locations and writing concepts, not only dates. But what did writers do with these histories, and how did they judge what the system showed them?

4:08 What writers did with the history

How did writers use the interfaces we just explored?

The history card names a formative study, which informed the design, and a longitudinal diary study, which followed writers over several sessions.

Writers used IRIS to locate specific revisions and understand how their writing progressed.

Writers compared system interpretations with their prior beliefs and personal experience, treating an output as something to consider rather than simply accept.

An interpretation could confirm an intuition, challenge it, or deepen understanding—and help guide future work.

The citation leads back to “Cited version V2312,” which writers could inspect and judge against the recorded changes and their own experience.

So the authors recommend keeping that evidence accessible. But how completely can these traces represent a writer’s thinking?

5:03 A useful trace is not the whole mind

IRIS uses writing history to help writers explore their process. But keystrokes are evidence about thinking, not recordings of thought, so an inferred account may differ from the writer’s experience.

Keystrokes provide clues about thinking, but not direct access to it.

The inferred account and the writer’s experience may agree or differ.

Even when it differs from the writer’s experience, an interpretation can still be useful.

The traces and model also miss the writer’s intentions, audience goals, and social context.

Participants wrote for about half an hour in the formative study and three and a half hours in the diary study.

Those periods may not capture the full lifecycle of long-form writing.

5:49 A privacy and ethics concern

The writing data and commercial language model are the two ends of this pathway. What risks come with that arrangement?

As writing data flows into the model, the authors raise privacy and ethics concerns.

One risk is that user data could be used for training without permission.

Another is that users may have limited remedies if their work’s rights are violated.

These are potential risks, not misuse established by the study.

IRIS makes writing history material for reflection, not an unquestionable account of thought, so writers can inspect and question its interpretations.